turbolinux.com (product: enfusion)
epaulson at students.wisc.edu
Thu Aug 10 12:40:58 PDT 2000
On Fri, 11 Aug 2000, Indraneel Majumdar wrote:
> Date: Fri, 11 Aug 2000 00:18:56 -0700 (PDT)
> To: David Lombard <david.lombard at mscsoftware.com>
> From: Indraneel Majumdar <indraneel at www.cdfd.org.in>
> Subject: Re: turbolinux.com (product: enfusion)
> I have some more queries which I have inserted in between the text.
> (Please remember that I seem to have less maths background than you are
> assuming, but I'm trying to understand) Thanks for your replies.
> On Thu, 10 Aug 2000, David Lombard wrote:
> > EnFuzion is a "design scan" program. That is, it runs tens, hundreds,
> > thousands, &etc of jobs to exhaustively search a design space. No
> > modifications are required to the program being run, as it's run under
> > control of a script.
> What exactly is a design scan? Is it the entire range of possible input
> and output values for an algorithm?
No. It's merely trying out as many different inputs as you've got the
time/resources/interest in trying.
> > But don't fool yourself into thinking you now have a "parallel" program
> > -- you don't.
> > What you do have is a tool for trying many combinations of input
> > parameters as quickly as your resources will allow. You could run
> > cases by hand (ouch) or with some scripts that use rsh(1) and rcp(1).
> > The typical lowest level process for a design scan is:
> > Specify the design variables and their ranges.
> How do I specify the variables and ranges without knowing the algorithm?
> Generally I'll have only one set of inputs. Do I have to run the program
> previously with different sets of inputs so that it can learn (similiar
> neural networks) the possible algorithm (or derive a simplified relation
> between input and output)?
You're expecting too much from this program. All it does is run your
many many times. If you had it run "Hello, World" a million times, it
would. Enfusion is not going to figure out that it's running "Hello, World"
and optimize itself for doing that.
> > Create a script that takes as input one value for each design variable,
> > runs the program, and extracts whatever results are appropriate.
> How does it know what results are appropriate?
> The program then also needs
> an error checking routine (reminds me of genetic algorithms).
That's totally beyond the scope of these sorts of systems.
More information about the Beowulf